Agents

A collection of Claude Code agents and skills for software development workflows — covering code review, planning, implementation, and continuous improvement.

Structure

agents/     Reusable sub-agents (launched by skills or directly)
skills/     Agent skills (/skill-name)

Agents

Agent	Description
code-reviewer	Multi-phase code review with severity scoring and confidence filtering
code-explorer	Traces execution paths and maps architecture to inform new development
coderabbit-reviewer	Code review via the CodeRabbit CLI
idea-stress-tester	Stress-tests feature ideas: surfaces risks, assumptions, and one-way door decisions

Skills

Skill	Description
commit	Creates conventional commits grouped by type (feat/fix/chore/…)
code-review	Runs the code-reviewer agent and walks through issues one at a time with user approval
pr	Commits changes, creates or links a Jira ticket, and opens a draft PR
plan-create	Explores the codebase, interviews the user, writes a phased implementation plan, stress-tests it, and breaks it into PRs
plan-implement	Picks the next pending PR from a plan, implements it, runs two rounds of agent review, then hands off to PR creation
plan-sync	Updates plan.md and pr-breakdown.md to match the latest codebase changes
interview	Interviews the user in depth about a spec and produces a finalized spec file
reflect	Analyzes the conversation for disagreements, proposes rule updates to AGENTS.md or skill files, and commits approved changes

Workflows

These two are all you need to know. Everything else runs behind the scenes.

1. /plan-create

/plan-create <feature description> to discuss with AI and come up with a plan. It will ask clarification questions to ensure it aligns with your expectations. At the end, it will generate a plan and a PR breakdown checklist to guide your implementation order.

The default plan mode works well for small-scope changes. This skill addresses that limitation by helping you think through an entire feature, not just a single PR.

Key highlights

1. It will run /interview to resolve any ambiguity. The AI may ask 40–50 questions until it is confident that all edge cases are covered.
2. It then uses idea-stress-tester to ensure the plan is solid. For example, verifying best practices and the use of up-to-date features or SDKs.
3. Unlike the default plan mode, this skill creates a PR breakdown to guide the implementation order, helping you avoid submitting massive 10k+ line PRs to your peers.

2. /plan-implement

For each PR, run /plan-implement <plan.md>, and it will handle the implementation and create a PR.

Key highlights

1. After completing the code changes, the AI runs two rounds of self-review before handing the work back to you. This is more robust than simply checking that the build, tests, and linter pass, and it’s more likely to follow the memory and rule files you’ve defined. It uses subagents to have isolated contexts and make the review more neutral. You can also spawn other subagents to call between codex and claude if needed.
2. Before creating the PR, it reflects on your feedback and updates the relevant memory, rule, and skill files to avoid making the same mistakes in the future.
3. When creating the PR, it will prompt you for the corresponding ticket in your issue tracking system (Jira in this case) and automatically update its progress, eliminating the need for manual syncing.

Setup

I believe everyone should customize their workflow to fit their needs, so copying/pasting those files to your repo will be better in long term.